Statistical Methods in Finance 2025

Data-Driven Early Warning Systems for Loan Default Prediction: Evidence from MSME-Linked Portfolios of a Public Sector Bank in India

By:Manish Dulal
Indian Institute of Technology, Bombay

"India’s economic growth and, by extension, the health of its public sector banks depend on the stability of the Micro, Small, and Medium Enterprises (MSME) sector. As a result, financial institutions and regulators have serious concerns about the prompt management of non-performing assets (NPAs). Despite extensive research on financial distress, research gaps remain for segment-specific, data-driven models that utilize authentic, bank-level data from emerging economies. To address this, the study aims to develop robust early warning systems for the timely identification of NPA across three borrower segments: the Mid-Segment (MS), Small Business Segment (SBS), and Small and Medium Enterprises (SME). Using verified internal financial and NPA data provided by the bank from 2022 to 2025, we construct 61 financial indicators and employ Logistic Regression, Random Forest, and XGBoost models, integrating SMOTE to address class imbalance to predict potential defaults. We find that advanced machine learning models consistently outperform linear models in prediction ac- curacy. The key predictors of default differ substantially across portfolios, underscoring that leverage and operating cost ratios are more significant drivers of distress than profitabil- ity in the SBS and SME segments. The key message is that a one-size-fits-all approach to credit risk modeling is inadequate; segment-specific models are empirically and operationally necessary. These contributions enhance risk management frameworks and policymaking in banking, supporting financial stability, and MSME sector growth in emerging markets."